Reducing Latency and Enhancing Accuracy in LLM Inference through Firmware-Level Optimization